[micheleRP]

Scheduled Connect sync against Connect v4.115.1 (current stable tag; changelog section dated 2026-10-23, nothing newer tagged). The benthos dependency is unchanged at v4.84.0 in go.mod at the tag, so no engine, CLI or Bloblang surface moved with the release.

The release itself

Nothing in the v4.115.1 changelog is a structural skill surface:

• The two kafka entries are fixes inside the franz writer — a data race between the produce-error and record-decoration slices, and the dispatch signal now firing when a record is handed to the producer, restoring the hand-off the ordered per-partition input relies on. No flag, field, component name or documented behaviour changes, and no skill asserted the broken behaviour.
• The oracledb_cdc entry is a schema-unit change, applied below.

The remaining changes are drift corrections and two durable CDC constraints the skills were missing. Per-field additions elsewhere in the window stay deferred to the generated reference and rpk connect create.

Changes

skills/connect/references/migration.md (Migration row of references/SOURCES.md)

Change	Verified in
The migrator input's lag gauge is redpanda_lag, not input_redpanda_migrator_lag (absent from the tree), and it is emitted only when the input has a consumer_group	redpanda-data/connect internal/impl/kafka/lag.go (lagMetricNameRedpanda), internal/impl/redpanda/migrator/migrator.go (both specs) and internal/impl/redpanda/migrator/README.md, at v4.115.1
A single migrator pair needs no labels; labels matter only with several pairs in one config, and a mismatched pair fails to connect rather than silently mis-coordinating. Labels dropped from the workflow YAML	internal/impl/redpanda/migrator/migrator.go (newMigratorFrom's migratorKey{label, stream} with its "default" fallback, the input spec's pairing paragraph, validateInitialized)
Label syntax replaced with the engine's own rule (lowercase [a-z0-9_-], no leading _/-); the previous "3–128 chars, A-Za-z0-9-_" claim was wrong on both counts	redpanda-data/benthos internal/docs/config.go (labelExpression, ValidateLabel) at v4.84.0
Exact consumer-group offset translation for groups in the Empty state via the output's offset header — the skill presented translation as timestamp-based and approximate, full stop	internal/impl/redpanda/migrator/migrator_groups.go (tryFindExactOffset, gated on State == "Empty" and a non-empty header) and the offset_header field spec in migrator.go
An existing destination topic with fewer partitions than its source has partitions added to match ("never overwritten" overstated it)	the output spec's guarantees in internal/impl/redpanda/migrator/migrator.go

skills/connect-cdc-postgres/ (SKILL.md, references/config-reference.md)

• postgres_cdc requires the primary instance — a read replica or physical standby is not supported. A product requirement stated in the connector's reference documentation; provenance is recorded in the skill's references/SOURCES.md.
• The snapshot phase holds the database's xmin horizon for as long as it runs, for every table, so a scan measured in hours defers cleanup of row versions that died after it started; large tables should use an incremental snapshot. Verified in internal/impl/postgresql/pglogicalstream/snapshotter.go (openTxn: read-only REPEATABLE READ on a background context, then SET TRANSACTION SNAPSHOT; rolled back only in releaseSnapshot) and pglogicalstream/logical_stream.go (SnapshotAction: "EXPORT_SNAPSHOT").

No postgres_cdc source changed between v4.115.0 and the tag.

skills/connect-cdc-oracle/references/pipeline-and-output.md

• TIMESTAMP columns declare microseconds in the schema the schema metadata field carries (DATE keeps milliseconds), so TIMESTAMP(6) is no longer truncated and the snapshot and streaming schemas agree; a consumer holding the pre-4.115.1 schema sees the unit change. Verified in internal/impl/oracledb/schema.go (columnToCommon, oracleTimestampLogical) and internal/impl/oracledb/TYPES.md at the tag. The JSON wire format of a timestamp is unchanged, so the type table still reads ISO-8601.

Checked and deliberately not documented

• The oracledb_cdc / microsoft_sql_server_cdc batcher refactor onto a shared flush-ticket package, and the second "flush queue is sealed" variant of their rebuild-publisher warning: internal, and no skill quotes those log lines.
• Durability: no field list, default, Bloblang catalog or component roster was pinned; the migrator's offset_header is named only where its semantics change a documented guarantee.

TODO for a human

Duplicate-label linting versus the documented multi-pair scheme. benthos v4.84.0 registers every component label in one per-run lint context and returns a duplicate-label lint for a label used twice in a config, which appears to collide with the migrator's documented scheme of giving an input and its paired output the same label. migration.md now tells readers to lint a multi-pair config and confirm the behaviour on their version rather than asserting either way. Worth a Connect-team answer on whether a deliberately matched pair lints clean.

---

🤖 Generated with Claude Code